sys/x86/acpica/srat.c

   1 /*-
   2  * Copyright (c) 2010 Advanced Computing Technologies LLC
   3  * Written by: John H. Baldwin <jhb@FreeBSD.org>
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 #include <sys/cdefs.h>
  29 __FBSDID("$FreeBSD$");
  30
  31 #include <sys/param.h>
  32 #include <sys/bus.h>
  33 #include <sys/kernel.h>
  34 #include <sys/smp.h>
  35 #include <vm/vm.h>
  36 #include <vm/pmap.h>
  37 #include <vm/vm_param.h>
  38 #include <vm/vm_phys.h>
  39
  40 #include <contrib/dev/acpica/include/acpi.h>
  41 #include <contrib/dev/acpica/include/actables.h>
  42
  43 #include <machine/intr_machdep.h>
  44 #include <machine/apicvar.h>
  45
  46 #include <dev/acpica/acpivar.h>
  47
  48 #if VM_NDOMAIN > 1
  49 struct cpu_info {
  50         int enabled:1;
  51         int has_memory:1;
  52         int domain;
  53 } cpus[MAX_APIC_ID + 1];
  54
  55 struct mem_affinity mem_info[VM_PHYSSEG_MAX + 1];
  56 int num_mem;
  57
  58 static ACPI_TABLE_SRAT *srat;
  59 static vm_paddr_t srat_physaddr;
  60
  61 static void     srat_walk_table(acpi_subtable_handler *handler, void *arg);
  62
  63 /*
  64  * Returns true if a memory range overlaps with at least one range in
  65  * phys_avail[].
  66  */
  67 static int
  68 overlaps_phys_avail(vm_paddr_t start, vm_paddr_t end)
  69 {
  70         int i;
  71
  72         for (i = 0; phys_avail[i] != 0 && phys_avail[i + 1] != 0; i += 2) {
  73                 if (phys_avail[i + 1] < start)
  74                         continue;
  75                 if (phys_avail[i] < end)
  76                         return (1);
  77                 break;
  78         }
  79         return (0);
  80
  81 }
  82
  83 static void
  84 srat_parse_entry(ACPI_SUBTABLE_HEADER *entry, void *arg)
  85 {
  86         ACPI_SRAT_CPU_AFFINITY *cpu;
  87         ACPI_SRAT_X2APIC_CPU_AFFINITY *x2apic;
  88         ACPI_SRAT_MEM_AFFINITY *mem;
  89         int domain, i, slot;
  90
  91         switch (entry->Type) {
  92         case ACPI_SRAT_TYPE_CPU_AFFINITY:
  93                 cpu = (ACPI_SRAT_CPU_AFFINITY *)entry;
  94                 domain = cpu->ProximityDomainLo |
  95                     cpu->ProximityDomainHi[0] << 8 |
  96                     cpu->ProximityDomainHi[1] << 16 |
  97                     cpu->ProximityDomainHi[2] << 24;
  98                 if (bootverbose)
  99                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 100                             cpu->ApicId, domain,
 101                             (cpu->Flags & ACPI_SRAT_CPU_ENABLED) ?
 102                             "enabled" : "disabled");
 103                 if (!(cpu->Flags & ACPI_SRAT_CPU_ENABLED))
 104                         break;
 105                 KASSERT(!cpus[cpu->ApicId].enabled,
 106                     ("Duplicate local APIC ID %u", cpu->ApicId));
 107                 cpus[cpu->ApicId].domain = domain;
 108                 cpus[cpu->ApicId].enabled = 1;
 109                 break;
 110         case ACPI_SRAT_TYPE_X2APIC_CPU_AFFINITY:
 111                 x2apic = (ACPI_SRAT_X2APIC_CPU_AFFINITY *)entry;
 112                 if (bootverbose)
 113                         printf("SRAT: Found CPU APIC ID %u domain %d: %s\n",
 114                             x2apic->ApicId, x2apic->ProximityDomain,
 115                             (x2apic->Flags & ACPI_SRAT_CPU_ENABLED) ?
 116                             "enabled" : "disabled");
 117                 if (!(x2apic->Flags & ACPI_SRAT_CPU_ENABLED))
 118                         break;
 119                 KASSERT(!cpus[x2apic->ApicId].enabled,
 120                     ("Duplicate local APIC ID %u", x2apic->ApicId));
 121                 cpus[x2apic->ApicId].domain = x2apic->ProximityDomain;
 122                 cpus[x2apic->ApicId].enabled = 1;
 123                 break;
 124         case ACPI_SRAT_TYPE_MEMORY_AFFINITY:
 125                 mem = (ACPI_SRAT_MEM_AFFINITY *)entry;
 126                 if (bootverbose)
 127                         printf(
 128                     "SRAT: Found memory domain %d addr %jx len %jx: %s\n",
 129                             mem->ProximityDomain, (uintmax_t)mem->BaseAddress,
 130                             (uintmax_t)mem->Length,
 131                             (mem->Flags & ACPI_SRAT_MEM_ENABLED) ?
 132                             "enabled" : "disabled");
 133                 if (!(mem->Flags & ACPI_SRAT_MEM_ENABLED))
 134                         break;
 135                 if (!overlaps_phys_avail(mem->BaseAddress,
 136                     mem->BaseAddress + mem->Length)) {
 137                         printf("SRAT: Ignoring memory at addr %jx\n",
 138                             (uintmax_t)mem->BaseAddress);
 139                         break;
 140                 }
 141                 if (num_mem == VM_PHYSSEG_MAX) {
 142                         printf("SRAT: Too many memory regions\n");
 143                         *(int *)arg = ENXIO;
 144                         break;
 145                 }
 146                 slot = num_mem;
 147                 for (i = 0; i < num_mem; i++) {
 148                         if (mem_info[i].end <= mem->BaseAddress)
 149                                 continue;
 150                         if (mem_info[i].start <
 151                             (mem->BaseAddress + mem->Length)) {
 152                                 printf("SRAT: Overlapping memory entries\n");
 153                                 *(int *)arg = ENXIO;
 154                                 return;
 155                         }
 156                         slot = i;
 157                 }
 158                 for (i = num_mem; i > slot; i--)
 159                         mem_info[i] = mem_info[i - 1];
 160                 mem_info[slot].start = mem->BaseAddress;
 161                 mem_info[slot].end = mem->BaseAddress + mem->Length;
 162                 mem_info[slot].domain = mem->ProximityDomain;
 163                 num_mem++;
 164                 break;
 165         }
 166 }
 167
 168 /*
 169  * Ensure each memory domain has at least one CPU and that each CPU
 170  * has at least one memory domain.
 171  */
 172 static int
 173 check_domains(void)
 174 {
 175         int found, i, j;
 176
 177         for (i = 0; i < num_mem; i++) {
 178                 found = 0;
 179                 for (j = 0; j <= MAX_APIC_ID; j++)
 180                         if (cpus[j].enabled &&
 181                             cpus[j].domain == mem_info[i].domain) {
 182                                 cpus[j].has_memory = 1;
 183                                 found++;
 184                         }
 185                 if (!found) {
 186                         printf("SRAT: No CPU found for memory domain %d\n",
 187                             mem_info[i].domain);
 188                         return (ENXIO);
 189                 }
 190         }
 191         for (i = 0; i <= MAX_APIC_ID; i++)
 192                 if (cpus[i].enabled && !cpus[i].has_memory) {
 193                         printf("SRAT: No memory found for CPU %d\n", i);
 194                         return (ENXIO);
 195                 }
 196         return (0);
 197 }
 198
 199 /*
 200  * Check that the SRAT memory regions cover all of the regions in
 201  * phys_avail[].
 202  */
 203 static int
 204 check_phys_avail(void)
 205 {
 206         vm_paddr_t address;
 207         int i, j;
 208
 209         /* j is the current offset into phys_avail[]. */
 210         address = phys_avail[0];
 211         j = 0;
 212         for (i = 0; i < num_mem; i++) {
 213                 /*
 214                  * Consume as many phys_avail[] entries as fit in this
 215                  * region.
 216                  */
 217                 while (address >= mem_info[i].start &&
 218                     address <= mem_info[i].end) {
 219                         /*
 220                          * If we cover the rest of this phys_avail[] entry,
 221                          * advance to the next entry.
 222                          */
 223                         if (phys_avail[j + 1] <= mem_info[i].end) {
 224                                 j += 2;
 225                                 if (phys_avail[j] == 0 &&
 226                                     phys_avail[j + 1] == 0) {
 227                                         return (0);
 228                                 }
 229                                 address = phys_avail[j];
 230                         } else
 231                                 address = mem_info[i].end + 1;
 232                 }
 233         }
 234         printf("SRAT: No memory region found for %jx - %jx\n",
 235             (uintmax_t)phys_avail[j], (uintmax_t)phys_avail[j + 1]);
 236         return (ENXIO);
 237 }
 238
 239 /*
 240  * Renumber the memory domains to be compact and zero-based if not
 241  * already.  Returns an error if there are too many domains.
 242  */
 243 static int
 244 renumber_domains(void)
 245 {
 246         int domains[VM_PHYSSEG_MAX];
 247         int ndomain, i, j, slot;
 248
 249         /* Enumerate all the domains. */
 250         ndomain = 0;
 251         for (i = 0; i < num_mem; i++) {
 252                 /* See if this domain is already known. */
 253                 for (j = 0; j < ndomain; j++) {
 254                         if (domains[j] >= mem_info[i].domain)
 255                                 break;
 256                 }
 257                 if (j < ndomain && domains[j] == mem_info[i].domain)
 258                         continue;
 259
 260                 /* Insert the new domain at slot 'j'. */
 261                 slot = j;
 262                 for (j = ndomain; j > slot; j--)
 263                         domains[j] = domains[j - 1];
 264                 domains[slot] = mem_info[i].domain;
 265                 ndomain++;
 266                 if (ndomain > VM_NDOMAIN) {
 267                         printf("SRAT: Too many memory domains\n");
 268                         return (EFBIG);
 269                 }
 270         }
 271
 272         /* Renumber each domain to its index in the sorted 'domains' list. */
 273         for (i = 0; i < ndomain; i++) {
 274                 /*
 275                  * If the domain is already the right value, no need
 276                  * to renumber.
 277                  */
 278                 if (domains[i] == i)
 279                         continue;
 280
 281                 /* Walk the cpu[] and mem_info[] arrays to renumber. */
 282                 for (j = 0; j < num_mem; j++)
 283                         if (mem_info[j].domain == domains[i])
 284                                 mem_info[j].domain = i;
 285                 for (j = 0; j <= MAX_APIC_ID; j++)
 286                         if (cpus[j].enabled && cpus[j].domain == domains[i])
 287                                 cpus[j].domain = i;
 288         }
 289         return (0);
 290 }
 291
 292 /*
 293  * Look for an ACPI System Resource Affinity Table ("SRAT")
 294  */
 295 static void
 296 parse_srat(void *dummy)
 297 {
 298         int error;
 299
 300         if (resource_disabled("srat", 0))
 301                 return;
 302
 303         srat_physaddr = acpi_find_table(ACPI_SIG_SRAT);
 304         if (srat_physaddr == 0)
 305                 return;
 306
 307         /*
 308          * Make a pass over the table to populate the cpus[] and
 309          * mem_info[] tables.
 310          */
 311         srat = acpi_map_table(srat_physaddr, ACPI_SIG_SRAT);
 312         error = 0;
 313         srat_walk_table(srat_parse_entry, &error);
 314         acpi_unmap_table(srat);
 315         srat = NULL;
 316         if (error || check_domains() != 0 || check_phys_avail() != 0 ||
 317             renumber_domains() != 0) {
 318                 srat_physaddr = 0;
 319                 return;
 320         }
 321
 322         /* Point vm_phys at our memory affinity table. */
 323         mem_affinity = mem_info;
 324 }
 325 SYSINIT(parse_srat, SI_SUB_VM - 1, SI_ORDER_FIRST, parse_srat, NULL);
 326
 327 static void
 328 srat_walk_table(acpi_subtable_handler *handler, void *arg)
 329 {
 330
 331         acpi_walk_subtables(srat + 1, (char *)srat + srat->Header.Length,
 332             handler, arg);
 333 }
 334
 335 /*
 336  * Setup per-CPU ACPI IDs.
 337  */
 338 static void
 339 srat_set_cpus(void *dummy)
 340 {
 341         struct cpu_info *cpu;
 342         struct pcpu *pc;
 343         u_int i;
 344
 345         if (srat_physaddr == 0)
 346                 return;
 347         for (i = 0; i < MAXCPU; i++) {
 348                 if (CPU_ABSENT(i))
 349                         continue;
 350                 pc = pcpu_find(i);
 351                 KASSERT(pc != NULL, ("no pcpu data for CPU %u", i));
 352                 cpu = &cpus[pc->pc_apic_id];
 353                 if (!cpu->enabled)
 354                         panic("SRAT: CPU with APIC ID %u is not known",
 355                             pc->pc_apic_id);
 356                 pc->pc_domain = cpu->domain;
 357                 if (bootverbose)
 358                         printf("SRAT: CPU %u has memory domain %d\n", i,
 359                             cpu->domain);
 360         }
 361 }
 362 SYSINIT(srat_set_cpus, SI_SUB_CPU, SI_ORDER_ANY, srat_set_cpus, NULL);
 363 #endif /* VM_NDOMAIN > 1 */